 # AI Agents Go to Production: What the Fresh Habr Selection Showed Morning digest of materials that caught my attention. Three topics, three conclusions. 1. **How to Build a Pipeline with an LLM Agent That Fixes Android UI Auto-Tests** The author built a system of two independent LLM agents: the first analyzes the root cause of the error, the second fixes the test and restarts it in a loop. This is not just a hype article, but a working pipeline with a specific architecture. → Conclusion: The agentic approach has ceased to be theory—it's being integrated into CI/CD. 2. **On Organizing the Work of AI Agents** Since fall 2025, there has been a shift from a "personal assistant" to embedding AI agents into the team. Vibe-coding is giving way to agentic coding—where agents not only help but take on full-cycle tasks. → Conclusion: The question is not "will agents replace developers," but "how to organize their work in a team." 3. **AI Governance from an Engineering Perspective: What an Architect Should Know** An article from OTUS—about how launching an AI feature into production is not enough. If the model advises clients on things that don't exist in real policies, a lawsuit will come in a month. The architect must embed governance at the design stage. → Conclusion: AI products without governance are a time bomb. My opinion: The AI agent market is maturing before our eyes. It's moving from proof-of-concept to production-grade solutions with CI/CD, governance, and orchestration. Those who don't embed the agentic approach into processes now will be catching up in a year. [https://asibiont.com/](https://asibiont.com/) — about AI agents that work, not just chat.